1#!/usr/bin/env perl
2
3# ====================================================================
4# [Re]written by Andy Polyakov <appro@openssl.org> for the OpenSSL
5# project. The module is, however, dual licensed under OpenSSL and
6# CRYPTOGAMS licenses depending on where you obtain it. For further
7# details see http://www.openssl.org/~appro/cryptogams/.
8# ====================================================================
9
10# "[Re]written" was achieved in two major overhauls. In 2004 BODY_*
11# functions were re-implemented to address P4 performance issue [see
12# commentary below], and in 2006 the rest was rewritten in order to
13# gain freedom to liberate licensing terms.
14
15# January, September 2004.
16#
17# It was noted that Intel IA-32 C compiler generates code which
18# performs ~30% *faster* on P4 CPU than original *hand-coded*
19# SHA1 assembler implementation. To address this problem (and
20# prove that humans are still better than machines:-), the
21# original code was overhauled, which resulted in following
22# performance changes:
23#
24#		compared with original	compared with Intel cc
25#		assembler impl.		generated code
26# Pentium	-16%			+48%
27# PIII/AMD	+8%			+16%
28# P4		+85%(!)			+45%
29#
30# As you can see Pentium came out as looser:-( Yet I reckoned that
31# improvement on P4 outweights the loss and incorporate this
32# re-tuned code to 0.9.7 and later.
33# ----------------------------------------------------------------
34#					<appro@fy.chalmers.se>
35
36# August 2009.
37#
38# George Spelvin has tipped that F_40_59(b,c,d) can be rewritten as
39# '(c&d) + (b&(c^d))', which allows to accumulate partial results
40# and lighten "pressure" on scratch registers. This resulted in
41# >12% performance improvement on contemporary AMD cores (with no
42# degradation on other CPUs:-). Also, the code was revised to maximize
43# "distance" between instructions producing input to 'lea' instruction
44# and the 'lea' instruction itself, which is essential for Intel Atom
45# core and resulted in ~15% improvement.
46
47# October 2010.
48#
49# Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it
50# is to offload message schedule denoted by Wt in NIST specification,
51# or Xupdate in OpenSSL source, to SIMD unit. The idea is not novel,
52# and in SSE2 context was first explored by Dean Gaudet in 2004, see
53# http://arctic.org/~dean/crypto/sha1.html. Since then several things
54# have changed that made it interesting again:
55#
56# a) XMM units became faster and wider;
57# b) instruction set became more versatile;
58# c) an important observation was made by Max Locktykhin, which made
59#    it possible to reduce amount of instructions required to perform
60#    the operation in question, for further details see
61#    http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/.
62
63# April 2011.
64#
65# Add AVX code path, probably most controversial... The thing is that
66# switch to AVX alone improves performance by as little as 4% in
67# comparison to SSSE3 code path. But below result doesn't look like
68# 4% improvement... Trouble is that Sandy Bridge decodes 'ro[rl]' as
69# pair of µ-ops, and it's the additional µ-ops, two per round, that
70# make it run slower than Core2 and Westmere. But 'sh[rl]d' is decoded
71# as single µ-op by Sandy Bridge and it's replacing 'ro[rl]' with
72# equivalent 'sh[rl]d' that is responsible for the impressive 5.1
73# cycles per processed byte. But 'sh[rl]d' is not something that used
74# to be fast, nor does it appear to be fast in upcoming Bulldozer
75# [according to its optimization manual]. Which is why AVX code path
76# is guarded by *both* AVX and synthetic bit denoting Intel CPUs.
77# One can argue that it's unfair to AMD, but without 'sh[rl]d' it
78# makes no sense to keep the AVX code path. If somebody feels that
79# strongly, it's probably more appropriate to discuss possibility of
80# using vector rotate XOP on AMD...
81
82# March 2014.
83#
84# Add support for Intel SHA Extensions.
85
86######################################################################
87# Current performance is summarized in following table. Numbers are
88# CPU clock cycles spent to process single byte (less is better).
89#
90#		x86		SSSE3		AVX
91# Pentium	15.7		-
92# PIII		11.5		-
93# P4		10.6		-
94# AMD K8	7.1		-
95# Core2		7.3		6.0/+22%	-
96# Westmere	7.3		5.5/+33%	-
97# Sandy Bridge	8.8		6.2/+40%	5.1(**)/+73%
98# Ivy Bridge	7.2		4.8/+51%	4.7(**)/+53%
99# Haswell	6.5		4.3/+51%	4.1(**)/+58%
100# Skylake	6.4		4.1/+55%	4.1(**)/+55%
101# Bulldozer	11.6		6.0/+92%
102# VIA Nano	10.6		7.5/+41%
103# Atom		12.5		9.3(*)/+35%
104# Silvermont	14.5		9.9(*)/+46%
105# Goldmont	8.8		6.7/+30%	1.7(***)/+415%
106#
107# (*)	Loop is 1056 instructions long and expected result is ~8.25.
108#	The discrepancy is because of front-end limitations, so
109#	called MS-ROM penalties, and on Silvermont even rotate's
110#	limited parallelism.
111#
112# (**)	As per above comment, the result is for AVX *plus* sh[rl]d.
113#
114# (***)	SHAEXT result
115
116$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
117push(@INC,"${dir}","${dir}../../perlasm");
118require "x86asm.pl";
119
120$output=pop;
121open STDOUT,">$output";
122
123&asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
124
125$xmm=$ymm=0;
126for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
127
128# In upstream, this is controlled by shelling out to the compiler to check
129# versions, but BoringSSL is intended to be used with pre-generated perlasm
130# output, so this isn't useful anyway.
131$ymm = 1;
132
133$ymm = 0 unless ($xmm);
134
135$shaext=$xmm;	### set to zero if compiling for 1.0.1
136
137# TODO(davidben): Consider enabling the Intel SHA Extensions code once it's
138# been tested.
139$shaext = 0;
140
141&external_label("OPENSSL_ia32cap_P") if ($xmm);
142
143
144$A="eax";
145$B="ebx";
146$C="ecx";
147$D="edx";
148$E="edi";
149$T="esi";
150$tmp1="ebp";
151
152@V=($A,$B,$C,$D,$E,$T);
153
154$alt=0;	# 1 denotes alternative IALU implementation, which performs
155	# 8% *worse* on P4, same on Westmere and Atom, 2% better on
156	# Sandy Bridge...
157
158sub BODY_00_15
159	{
160	local($n,$a,$b,$c,$d,$e,$f)=@_;
161
162	&comment("00_15 $n");
163
164	&mov($f,$c);			# f to hold F_00_19(b,c,d)
165	 if ($n==0)  { &mov($tmp1,$a); }
166	 else        { &mov($a,$tmp1); }
167	&rotl($tmp1,5);			# tmp1=ROTATE(a,5)
168	 &xor($f,$d);
169	&add($tmp1,$e);			# tmp1+=e;
170	 &mov($e,&swtmp($n%16));	# e becomes volatile and is loaded
171	 				# with xi, also note that e becomes
172					# f in next round...
173	&and($f,$b);
174	&rotr($b,2);			# b=ROTATE(b,30)
175	 &xor($f,$d);			# f holds F_00_19(b,c,d)
176	&lea($tmp1,&DWP(0x5a827999,$tmp1,$e));	# tmp1+=K_00_19+xi
177
178	if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round
179		      &add($f,$tmp1); }	# f+=tmp1
180	else        { &add($tmp1,$f); }	# f becomes a in next round
181	&mov($tmp1,$a)			if ($alt && $n==15);
182	}
183
184sub BODY_16_19
185	{
186	local($n,$a,$b,$c,$d,$e,$f)=@_;
187
188	&comment("16_19 $n");
189
190if ($alt) {
191	&xor($c,$d);
192	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
193	&and($tmp1,$c);			# tmp1 to hold F_00_19(b,c,d), b&=c^d
194	 &xor($f,&swtmp(($n+8)%16));
195	&xor($tmp1,$d);			# tmp1=F_00_19(b,c,d)
196	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
197	&rotl($f,1);			# f=ROTATE(f,1)
198	 &add($e,$tmp1);		# e+=F_00_19(b,c,d)
199	&xor($c,$d);			# restore $c
200	 &mov($tmp1,$a);		# b in next round
201	&rotr($b,$n==16?2:7);		# b=ROTATE(b,30)
202	 &mov(&swtmp($n%16),$f);	# xi=f
203	&rotl($a,5);			# ROTATE(a,5)
204	 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
205	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
206	 &add($f,$a);			# f+=ROTATE(a,5)
207} else {
208	&mov($tmp1,$c);			# tmp1 to hold F_00_19(b,c,d)
209	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
210	&xor($tmp1,$d);
211	 &xor($f,&swtmp(($n+8)%16));
212	&and($tmp1,$b);
213	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
214	&rotl($f,1);			# f=ROTATE(f,1)
215	 &xor($tmp1,$d);		# tmp1=F_00_19(b,c,d)
216	&add($e,$tmp1);			# e+=F_00_19(b,c,d)
217	 &mov($tmp1,$a);
218	&rotr($b,2);			# b=ROTATE(b,30)
219	 &mov(&swtmp($n%16),$f);	# xi=f
220	&rotl($tmp1,5);			# ROTATE(a,5)
221	 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
222	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
223	 &add($f,$tmp1);		# f+=ROTATE(a,5)
224}
225	}
226
227sub BODY_20_39
228	{
229	local($n,$a,$b,$c,$d,$e,$f)=@_;
230	local $K=($n<40)?0x6ed9eba1:0xca62c1d6;
231
232	&comment("20_39 $n");
233
234if ($alt) {
235	&xor($tmp1,$c);			# tmp1 to hold F_20_39(b,c,d), b^=c
236	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
237	&xor($tmp1,$d);			# tmp1 holds F_20_39(b,c,d)
238	 &xor($f,&swtmp(($n+8)%16));
239	&add($e,$tmp1);			# e+=F_20_39(b,c,d)
240	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
241	&rotl($f,1);			# f=ROTATE(f,1)
242	 &mov($tmp1,$a);		# b in next round
243	&rotr($b,7);			# b=ROTATE(b,30)
244	 &mov(&swtmp($n%16),$f)		if($n<77);# xi=f
245	&rotl($a,5);			# ROTATE(a,5)
246	 &xor($b,$c)			if($n==39);# warm up for BODY_40_59
247	&and($tmp1,$b)			if($n==39);
248	 &lea($f,&DWP($K,$f,$e));	# f+=e+K_XX_YY
249	&mov($e,&swtmp(($n+1)%16))	if($n<79);# pre-fetch f for next round
250	 &add($f,$a);			# f+=ROTATE(a,5)
251	&rotr($a,5)			if ($n==79);
252} else {
253	&mov($tmp1,$b);			# tmp1 to hold F_20_39(b,c,d)
254	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
255	&xor($tmp1,$c);
256	 &xor($f,&swtmp(($n+8)%16));
257	&xor($tmp1,$d);			# tmp1 holds F_20_39(b,c,d)
258	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
259	&rotl($f,1);			# f=ROTATE(f,1)
260	 &add($e,$tmp1);		# e+=F_20_39(b,c,d)
261	&rotr($b,2);			# b=ROTATE(b,30)
262	 &mov($tmp1,$a);
263	&rotl($tmp1,5);			# ROTATE(a,5)
264	 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
265	&lea($f,&DWP($K,$f,$e));	# f+=e+K_XX_YY
266	 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
267	&add($f,$tmp1);			# f+=ROTATE(a,5)
268}
269	}
270
271sub BODY_40_59
272	{
273	local($n,$a,$b,$c,$d,$e,$f)=@_;
274
275	&comment("40_59 $n");
276
277if ($alt) {
278	&add($e,$tmp1);			# e+=b&(c^d)
279	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
280	&mov($tmp1,$d);
281	 &xor($f,&swtmp(($n+8)%16));
282	&xor($c,$d);			# restore $c
283	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
284	&rotl($f,1);			# f=ROTATE(f,1)
285	 &and($tmp1,$c);
286	&rotr($b,7);			# b=ROTATE(b,30)
287	 &add($e,$tmp1);		# e+=c&d
288	&mov($tmp1,$a);			# b in next round
289	 &mov(&swtmp($n%16),$f);	# xi=f
290	&rotl($a,5);			# ROTATE(a,5)
291	 &xor($b,$c)			if ($n<59);
292	&and($tmp1,$b)			if ($n<59);# tmp1 to hold F_40_59(b,c,d)
293	 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d))
294	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
295	 &add($f,$a);			# f+=ROTATE(a,5)
296} else {
297	&mov($tmp1,$c);			# tmp1 to hold F_40_59(b,c,d)
298	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
299	&xor($tmp1,$d);
300	 &xor($f,&swtmp(($n+8)%16));
301	&and($tmp1,$b);
302	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
303	&rotl($f,1);			# f=ROTATE(f,1)
304	 &add($tmp1,$e);		# b&(c^d)+=e
305	&rotr($b,2);			# b=ROTATE(b,30)
306	 &mov($e,$a);			# e becomes volatile
307	&rotl($e,5);			# ROTATE(a,5)
308	 &mov(&swtmp($n%16),$f);	# xi=f
309	&lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d))
310	 &mov($tmp1,$c);
311	&add($f,$e);			# f+=ROTATE(a,5)
312	 &and($tmp1,$d);
313	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
314	 &add($f,$tmp1);		# f+=c&d
315}
316	}
317
318&function_begin("sha1_block_data_order");
319if ($xmm) {
320  &static_label("shaext_shortcut")	if ($shaext);
321  &static_label("ssse3_shortcut");
322  &static_label("avx_shortcut")		if ($ymm);
323  &static_label("K_XX_XX");
324
325	&call	(&label("pic_point"));	# make it PIC!
326  &set_label("pic_point");
327	&blindpop($tmp1);
328	&picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point"));
329	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
330
331	&mov	($A,&DWP(0,$T));
332	&mov	($D,&DWP(4,$T));
333	&test	($D,1<<9);		# check SSSE3 bit
334	&jz	(&label("x86"));
335	&mov	($C,&DWP(8,$T));
336	&test	($A,1<<24);		# check FXSR bit
337	&jz	(&label("x86"));
338	if ($shaext) {
339		&test	($C,1<<29);		# check SHA bit
340		&jnz	(&label("shaext_shortcut"));
341	}
342	if ($ymm) {
343		&and	($D,1<<28);		# mask AVX bit
344		&and	($A,1<<30);		# mask "Intel CPU" bit
345		&or	($A,$D);
346		&cmp	($A,1<<28|1<<30);
347		&je	(&label("avx_shortcut"));
348	}
349	&jmp	(&label("ssse3_shortcut"));
350  &set_label("x86",16);
351}
352	&mov($tmp1,&wparam(0));	# SHA_CTX *c
353	&mov($T,&wparam(1));	# const void *input
354	&mov($A,&wparam(2));	# size_t num
355	&stack_push(16+3);	# allocate X[16]
356	&shl($A,6);
357	&add($A,$T);
358	&mov(&wparam(2),$A);	# pointer beyond the end of input
359	&mov($E,&DWP(16,$tmp1));# pre-load E
360	&jmp(&label("loop"));
361
362&set_label("loop",16);
363
364	# copy input chunk to X, but reversing byte order!
365	for ($i=0; $i<16; $i+=4)
366		{
367		&mov($A,&DWP(4*($i+0),$T));
368		&mov($B,&DWP(4*($i+1),$T));
369		&mov($C,&DWP(4*($i+2),$T));
370		&mov($D,&DWP(4*($i+3),$T));
371		&bswap($A);
372		&bswap($B);
373		&bswap($C);
374		&bswap($D);
375		&mov(&swtmp($i+0),$A);
376		&mov(&swtmp($i+1),$B);
377		&mov(&swtmp($i+2),$C);
378		&mov(&swtmp($i+3),$D);
379		}
380	&mov(&wparam(1),$T);	# redundant in 1st spin
381
382	&mov($A,&DWP(0,$tmp1));	# load SHA_CTX
383	&mov($B,&DWP(4,$tmp1));
384	&mov($C,&DWP(8,$tmp1));
385	&mov($D,&DWP(12,$tmp1));
386	# E is pre-loaded
387
388	for($i=0;$i<16;$i++)	{ &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
389	for(;$i<20;$i++)	{ &BODY_16_19($i,@V); unshift(@V,pop(@V)); }
390	for(;$i<40;$i++)	{ &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
391	for(;$i<60;$i++)	{ &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
392	for(;$i<80;$i++)	{ &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
393
394	(($V[5] eq $D) and ($V[0] eq $E)) or die;	# double-check
395
396	&mov($tmp1,&wparam(0));	# re-load SHA_CTX*
397	&mov($D,&wparam(1));	# D is last "T" and is discarded
398
399	&add($E,&DWP(0,$tmp1));	# E is last "A"...
400	&add($T,&DWP(4,$tmp1));
401	&add($A,&DWP(8,$tmp1));
402	&add($B,&DWP(12,$tmp1));
403	&add($C,&DWP(16,$tmp1));
404
405	&mov(&DWP(0,$tmp1),$E);	# update SHA_CTX
406	 &add($D,64);		# advance input pointer
407	&mov(&DWP(4,$tmp1),$T);
408	 &cmp($D,&wparam(2));	# have we reached the end yet?
409	&mov(&DWP(8,$tmp1),$A);
410	 &mov($E,$C);		# C is last "E" which needs to be "pre-loaded"
411	&mov(&DWP(12,$tmp1),$B);
412	 &mov($T,$D);		# input pointer
413	&mov(&DWP(16,$tmp1),$C);
414	&jb(&label("loop"));
415
416	&stack_pop(16+3);
417&function_end("sha1_block_data_order");
418
419if ($xmm) {
420if ($shaext) {
421######################################################################
422# Intel SHA Extensions implementation of SHA1 update function.
423#
424my ($ctx,$inp,$num)=("edi","esi","ecx");
425my ($ABCD,$E,$E_,$BSWAP)=map("xmm$_",(0..3));
426my @MSG=map("xmm$_",(4..7));
427
428sub sha1rnds4 {
429 my ($dst,$src,$imm)=@_;
430    if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
431    {	&data_byte(0x0f,0x3a,0xcc,0xc0|($1<<3)|$2,$imm);	}
432}
433sub sha1op38 {
434 my ($opcodelet,$dst,$src)=@_;
435    if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
436    {	&data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2);	}
437}
438sub sha1nexte	{ sha1op38(0xc8,@_); }
439sub sha1msg1	{ sha1op38(0xc9,@_); }
440sub sha1msg2	{ sha1op38(0xca,@_); }
441
442&function_begin("_sha1_block_data_order_shaext");
443	&call	(&label("pic_point"));	# make it PIC!
444	&set_label("pic_point");
445	&blindpop($tmp1);
446	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
447&set_label("shaext_shortcut");
448	&mov	($ctx,&wparam(0));
449	&mov	("ebx","esp");
450	&mov	($inp,&wparam(1));
451	&mov	($num,&wparam(2));
452	&sub	("esp",32);
453
454	&movdqu	($ABCD,&QWP(0,$ctx));
455	&movd	($E,&DWP(16,$ctx));
456	&and	("esp",-32);
457	&movdqa	($BSWAP,&QWP(0x50,$tmp1));	# byte-n-word swap
458
459	&movdqu	(@MSG[0],&QWP(0,$inp));
460	&pshufd	($ABCD,$ABCD,0b00011011);	# flip word order
461	&movdqu	(@MSG[1],&QWP(0x10,$inp));
462	&pshufd	($E,$E,0b00011011);		# flip word order
463	&movdqu	(@MSG[2],&QWP(0x20,$inp));
464	&pshufb	(@MSG[0],$BSWAP);
465	&movdqu	(@MSG[3],&QWP(0x30,$inp));
466	&pshufb	(@MSG[1],$BSWAP);
467	&pshufb	(@MSG[2],$BSWAP);
468	&pshufb	(@MSG[3],$BSWAP);
469	&jmp	(&label("loop_shaext"));
470
471&set_label("loop_shaext",16);
472	&dec		($num);
473	&lea		("eax",&DWP(0x40,$inp));
474	&movdqa		(&QWP(0,"esp"),$E);	# offload $E
475	&paddd		($E,@MSG[0]);
476	&cmovne		($inp,"eax");
477	&movdqa		(&QWP(16,"esp"),$ABCD);	# offload $ABCD
478
479for($i=0;$i<20-4;$i+=2) {
480	&sha1msg1	(@MSG[0],@MSG[1]);
481	&movdqa		($E_,$ABCD);
482	&sha1rnds4	($ABCD,$E,int($i/5));	# 0-3...
483	&sha1nexte	($E_,@MSG[1]);
484	&pxor		(@MSG[0],@MSG[2]);
485	&sha1msg1	(@MSG[1],@MSG[2]);
486	&sha1msg2	(@MSG[0],@MSG[3]);
487
488	&movdqa		($E,$ABCD);
489	&sha1rnds4	($ABCD,$E_,int(($i+1)/5));
490	&sha1nexte	($E,@MSG[2]);
491	&pxor		(@MSG[1],@MSG[3]);
492	&sha1msg2	(@MSG[1],@MSG[0]);
493
494	push(@MSG,shift(@MSG));	push(@MSG,shift(@MSG));
495}
496	&movdqu		(@MSG[0],&QWP(0,$inp));
497	&movdqa		($E_,$ABCD);
498	&sha1rnds4	($ABCD,$E,3);		# 64-67
499	&sha1nexte	($E_,@MSG[1]);
500	&movdqu		(@MSG[1],&QWP(0x10,$inp));
501	&pshufb		(@MSG[0],$BSWAP);
502
503	&movdqa		($E,$ABCD);
504	&sha1rnds4	($ABCD,$E_,3);		# 68-71
505	&sha1nexte	($E,@MSG[2]);
506	&movdqu		(@MSG[2],&QWP(0x20,$inp));
507	&pshufb		(@MSG[1],$BSWAP);
508
509	&movdqa		($E_,$ABCD);
510	&sha1rnds4	($ABCD,$E,3);		# 72-75
511	&sha1nexte	($E_,@MSG[3]);
512	&movdqu		(@MSG[3],&QWP(0x30,$inp));
513	&pshufb		(@MSG[2],$BSWAP);
514
515	&movdqa		($E,$ABCD);
516	&sha1rnds4	($ABCD,$E_,3);		# 76-79
517	&movdqa		($E_,&QWP(0,"esp"));
518	&pshufb		(@MSG[3],$BSWAP);
519	&sha1nexte	($E,$E_);
520	&paddd		($ABCD,&QWP(16,"esp"));
521
522	&jnz		(&label("loop_shaext"));
523
524	&pshufd	($ABCD,$ABCD,0b00011011);
525	&pshufd	($E,$E,0b00011011);
526	&movdqu	(&QWP(0,$ctx),$ABCD)
527	&movd	(&DWP(16,$ctx),$E);
528	&mov	("esp","ebx");
529&function_end("_sha1_block_data_order_shaext");
530}
531######################################################################
532# The SSSE3 implementation.
533#
534# %xmm[0-7] are used as ring @X[] buffer containing quadruples of last
535# 32 elements of the message schedule or Xupdate outputs. First 4
536# quadruples are simply byte-swapped input, next 4 are calculated
537# according to method originally suggested by Dean Gaudet (modulo
538# being implemented in SSSE3). Once 8 quadruples or 32 elements are
539# collected, it switches to routine proposed by Max Locktyukhin.
540#
541# Calculations inevitably require temporary reqisters, and there are
542# no %xmm registers left to spare. For this reason part of the ring
543# buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring
544# buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] -
545# X[-5], and X[4] - X[-4]...
546#
547# Another notable optimization is aggressive stack frame compression
548# aiming to minimize amount of 9-byte instructions...
549#
550# Yet another notable optimization is "jumping" $B variable. It means
551# that there is no register permanently allocated for $B value. This
552# allowed to eliminate one instruction from body_20_39...
553#
554my $Xi=4;			# 4xSIMD Xupdate round, start pre-seeded
555my @X=map("xmm$_",(4..7,0..3));	# pre-seeded for $Xi=4
556my @V=($A,$B,$C,$D,$E);
557my $j=0;			# hash round
558my $rx=0;
559my @T=($T,$tmp1);
560my $inp;
561
562my $_rol=sub { &rol(@_) };
563my $_ror=sub { &ror(@_) };
564
565&function_begin("_sha1_block_data_order_ssse3");
566	&call	(&label("pic_point"));	# make it PIC!
567	&set_label("pic_point");
568	&blindpop($tmp1);
569	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
570&set_label("ssse3_shortcut");
571
572	&movdqa	(@X[3],&QWP(0,$tmp1));		# K_00_19
573	&movdqa	(@X[4],&QWP(16,$tmp1));		# K_20_39
574	&movdqa	(@X[5],&QWP(32,$tmp1));		# K_40_59
575	&movdqa	(@X[6],&QWP(48,$tmp1));		# K_60_79
576	&movdqa	(@X[2],&QWP(64,$tmp1));		# pbswap mask
577
578	&mov	($E,&wparam(0));		# load argument block
579	&mov	($inp=@T[1],&wparam(1));
580	&mov	($D,&wparam(2));
581	&mov	(@T[0],"esp");
582
583	# stack frame layout
584	#
585	# +0	X[0]+K	X[1]+K	X[2]+K	X[3]+K	# XMM->IALU xfer area
586	#	X[4]+K	X[5]+K	X[6]+K	X[7]+K
587	#	X[8]+K	X[9]+K	X[10]+K	X[11]+K
588	#	X[12]+K	X[13]+K	X[14]+K	X[15]+K
589	#
590	# +64	X[0]	X[1]	X[2]	X[3]	# XMM->XMM backtrace area
591	#	X[4]	X[5]	X[6]	X[7]
592	#	X[8]	X[9]	X[10]	X[11]	# even borrowed for K_00_19
593	#
594	# +112	K_20_39	K_20_39	K_20_39	K_20_39	# constants
595	#	K_40_59	K_40_59	K_40_59	K_40_59
596	#	K_60_79	K_60_79	K_60_79	K_60_79
597	#	K_00_19	K_00_19	K_00_19	K_00_19
598	#	pbswap mask
599	#
600	# +192	ctx				# argument block
601	# +196	inp
602	# +200	end
603	# +204	esp
604	&sub	("esp",208);
605	&and	("esp",-64);
606
607	&movdqa	(&QWP(112+0,"esp"),@X[4]);	# copy constants
608	&movdqa	(&QWP(112+16,"esp"),@X[5]);
609	&movdqa	(&QWP(112+32,"esp"),@X[6]);
610	&shl	($D,6);				# len*64
611	&movdqa	(&QWP(112+48,"esp"),@X[3]);
612	&add	($D,$inp);			# end of input
613	&movdqa	(&QWP(112+64,"esp"),@X[2]);
614	&add	($inp,64);
615	&mov	(&DWP(192+0,"esp"),$E);		# save argument block
616	&mov	(&DWP(192+4,"esp"),$inp);
617	&mov	(&DWP(192+8,"esp"),$D);
618	&mov	(&DWP(192+12,"esp"),@T[0]);	# save original %esp
619
620	&mov	($A,&DWP(0,$E));		# load context
621	&mov	($B,&DWP(4,$E));
622	&mov	($C,&DWP(8,$E));
623	&mov	($D,&DWP(12,$E));
624	&mov	($E,&DWP(16,$E));
625	&mov	(@T[0],$B);			# magic seed
626
627	&movdqu	(@X[-4&7],&QWP(-64,$inp));	# load input to %xmm[0-3]
628	&movdqu	(@X[-3&7],&QWP(-48,$inp));
629	&movdqu	(@X[-2&7],&QWP(-32,$inp));
630	&movdqu	(@X[-1&7],&QWP(-16,$inp));
631	&pshufb	(@X[-4&7],@X[2]);		# byte swap
632	&pshufb	(@X[-3&7],@X[2]);
633	&pshufb	(@X[-2&7],@X[2]);
634	&movdqa	(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
635	&pshufb	(@X[-1&7],@X[2]);
636	&paddd	(@X[-4&7],@X[3]);		# add K_00_19
637	&paddd	(@X[-3&7],@X[3]);
638	&paddd	(@X[-2&7],@X[3]);
639	&movdqa	(&QWP(0,"esp"),@X[-4&7]);	# X[]+K xfer to IALU
640	&psubd	(@X[-4&7],@X[3]);		# restore X[]
641	&movdqa	(&QWP(0+16,"esp"),@X[-3&7]);
642	&psubd	(@X[-3&7],@X[3]);
643	&movdqa	(&QWP(0+32,"esp"),@X[-2&7]);
644	&mov	(@T[1],$C);
645	&psubd	(@X[-2&7],@X[3]);
646	&xor	(@T[1],$D);
647	&pshufd	(@X[0],@X[-4&7],0xee);		# was &movdqa	(@X[0],@X[-3&7]);
648	&and	(@T[0],@T[1]);
649	&jmp	(&label("loop"));
650
651######################################################################
652# SSE instruction sequence is first broken to groups of indepentent
653# instructions, independent in respect to their inputs and shifter
654# (not all architectures have more than one). Then IALU instructions
655# are "knitted in" between the SSE groups. Distance is maintained for
656# SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer
657# [which allegedly also implements SSSE3]...
658#
659# Temporary registers usage. X[2] is volatile at the entry and at the
660# end is restored from backtrace ring buffer. X[3] is expected to
661# contain current K_XX_XX constant and is used to caclulate X[-1]+K
662# from previous round, it becomes volatile the moment the value is
663# saved to stack for transfer to IALU. X[4] becomes volatile whenever
664# X[-4] is accumulated and offloaded to backtrace ring buffer, at the
665# end it is loaded with next K_XX_XX [which becomes X[3] in next
666# round]...
667#
668sub Xupdate_ssse3_16_31()		# recall that $Xi starts wtih 4
669{ use integer;
670  my $body = shift;
671  my @insns = (&$body,&$body,&$body,&$body);	# 40 instructions
672  my ($a,$b,$c,$d,$e);
673
674	 eval(shift(@insns));		# ror
675	 eval(shift(@insns));
676	 eval(shift(@insns));
677	&punpcklqdq(@X[0],@X[-3&7]);	# compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
678	&movdqa	(@X[2],@X[-1&7]);
679	 eval(shift(@insns));
680	 eval(shift(@insns));
681
682	  &paddd	(@X[3],@X[-1&7]);
683	  &movdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
684	 eval(shift(@insns));		# rol
685	 eval(shift(@insns));
686	&psrldq	(@X[2],4);		# "X[-3]", 3 dwords
687	 eval(shift(@insns));
688	 eval(shift(@insns));
689	&pxor	(@X[0],@X[-4&7]);	# "X[0]"^="X[-16]"
690	 eval(shift(@insns));
691	 eval(shift(@insns));		# ror
692
693	&pxor	(@X[2],@X[-2&7]);	# "X[-3]"^"X[-8]"
694	 eval(shift(@insns));
695	 eval(shift(@insns));
696	 eval(shift(@insns));
697
698	&pxor	(@X[0],@X[2]);		# "X[0]"^="X[-3]"^"X[-8]"
699	 eval(shift(@insns));
700	 eval(shift(@insns));		# rol
701	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
702	 eval(shift(@insns));
703	 eval(shift(@insns));
704
705	&movdqa	(@X[4],@X[0]);
706	 eval(shift(@insns));
707	 eval(shift(@insns));
708	 eval(shift(@insns));		# ror
709	&movdqa (@X[2],@X[0]);
710	 eval(shift(@insns));
711
712	&pslldq	(@X[4],12);		# "X[0]"<<96, extract one dword
713	&paddd	(@X[0],@X[0]);
714	 eval(shift(@insns));
715	 eval(shift(@insns));
716
717	&psrld	(@X[2],31);
718	 eval(shift(@insns));
719	 eval(shift(@insns));		# rol
720	&movdqa	(@X[3],@X[4]);
721	 eval(shift(@insns));
722	 eval(shift(@insns));
723	 eval(shift(@insns));
724
725	&psrld	(@X[4],30);
726	 eval(shift(@insns));
727	 eval(shift(@insns));		# ror
728	&por	(@X[0],@X[2]);		# "X[0]"<<<=1
729	 eval(shift(@insns));
730	  &movdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
731	 eval(shift(@insns));
732	 eval(shift(@insns));
733
734	&pslld	(@X[3],2);
735	 eval(shift(@insns));
736	 eval(shift(@insns));		# rol
737	&pxor   (@X[0],@X[4]);
738	  &movdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
739	 eval(shift(@insns));
740	 eval(shift(@insns));
741
742	&pxor	(@X[0],@X[3]);		# "X[0]"^=("X[0]"<<96)<<<2
743	  &pshufd	(@X[1],@X[-3&7],0xee)	if ($Xi<7);	# was &movdqa	(@X[1],@X[-2&7])
744	  &pshufd	(@X[3],@X[-1&7],0xee)	if ($Xi==7);
745	 eval(shift(@insns));
746	 eval(shift(@insns));
747
748	 foreach (@insns) { eval; }	# remaining instructions [if any]
749
750  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
751}
752
753sub Xupdate_ssse3_32_79()
754{ use integer;
755  my $body = shift;
756  my @insns = (&$body,&$body,&$body,&$body);	# 32 to 44 instructions
757  my ($a,$b,$c,$d,$e);
758
759	 eval(shift(@insns));		# body_20_39
760	&pxor	(@X[0],@X[-4&7]);	# "X[0]"="X[-32]"^"X[-16]"
761	&punpcklqdq(@X[2],@X[-1&7]);	# compose "X[-6]", was &palignr(@X[2],@X[-2&7],8)
762	 eval(shift(@insns));
763	 eval(shift(@insns));
764	 eval(shift(@insns));		# rol
765
766	&pxor	(@X[0],@X[-7&7]);	# "X[0]"^="X[-28]"
767	  &movdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
768	 eval(shift(@insns));
769	 eval(shift(@insns));
770	 eval(shift(@insns))		if (@insns[0] =~ /_rol/);
771	 if ($Xi%5) {
772	  &movdqa	(@X[4],@X[3]);	# "perpetuate" K_XX_XX...
773	 } else {			# ... or load next one
774	  &movdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
775	 }
776	 eval(shift(@insns));		# ror
777	  &paddd	(@X[3],@X[-1&7]);
778	 eval(shift(@insns));
779
780	&pxor	(@X[0],@X[2]);		# "X[0]"^="X[-6]"
781	 eval(shift(@insns));		# body_20_39
782	 eval(shift(@insns));
783	 eval(shift(@insns));
784	 eval(shift(@insns));		# rol
785
786	&movdqa	(@X[2],@X[0]);
787	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
788	 eval(shift(@insns));
789	 eval(shift(@insns));
790	 eval(shift(@insns));		# ror
791	 eval(shift(@insns));
792	 eval(shift(@insns))		if (@insns[0] =~ /_rol/);
793
794	&pslld	(@X[0],2);
795	 eval(shift(@insns));		# body_20_39
796	 eval(shift(@insns));
797	&psrld	(@X[2],30);
798	 eval(shift(@insns));
799	 eval(shift(@insns));		# rol
800	 eval(shift(@insns));
801	 eval(shift(@insns));
802	 eval(shift(@insns));		# ror
803	 eval(shift(@insns));
804	 eval(shift(@insns))		if (@insns[1] =~ /_rol/);
805	 eval(shift(@insns))		if (@insns[0] =~ /_rol/);
806
807	&por	(@X[0],@X[2]);		# "X[0]"<<<=2
808	 eval(shift(@insns));		# body_20_39
809	 eval(shift(@insns));
810	  &movdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
811	 eval(shift(@insns));
812	 eval(shift(@insns));		# rol
813	 eval(shift(@insns));
814	 eval(shift(@insns));
815	 eval(shift(@insns));		# ror
816	  &pshufd	(@X[3],@X[-1],0xee)	if ($Xi<19);	# was &movdqa	(@X[3],@X[0])
817	 eval(shift(@insns));
818
819	 foreach (@insns) { eval; }	# remaining instructions
820
821  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
822}
823
824sub Xuplast_ssse3_80()
825{ use integer;
826  my $body = shift;
827  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
828  my ($a,$b,$c,$d,$e);
829
830	 eval(shift(@insns));
831	 eval(shift(@insns));
832	 eval(shift(@insns));
833	 eval(shift(@insns));
834	 eval(shift(@insns));
835	 eval(shift(@insns));
836	 eval(shift(@insns));
837	  &paddd	(@X[3],@X[-1&7]);
838	 eval(shift(@insns));
839	 eval(shift(@insns));
840	 eval(shift(@insns));
841	 eval(shift(@insns));
842
843	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU
844
845	 foreach (@insns) { eval; }		# remaining instructions
846
847	&mov	($inp=@T[1],&DWP(192+4,"esp"));
848	&cmp	($inp,&DWP(192+8,"esp"));
849	&je	(&label("done"));
850
851	&movdqa	(@X[3],&QWP(112+48,"esp"));	# K_00_19
852	&movdqa	(@X[2],&QWP(112+64,"esp"));	# pbswap mask
853	&movdqu	(@X[-4&7],&QWP(0,$inp));	# load input
854	&movdqu	(@X[-3&7],&QWP(16,$inp));
855	&movdqu	(@X[-2&7],&QWP(32,$inp));
856	&movdqu	(@X[-1&7],&QWP(48,$inp));
857	&add	($inp,64);
858	&pshufb	(@X[-4&7],@X[2]);		# byte swap
859	&mov	(&DWP(192+4,"esp"),$inp);
860	&movdqa	(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
861
862  $Xi=0;
863}
864
865sub Xloop_ssse3()
866{ use integer;
867  my $body = shift;
868  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
869  my ($a,$b,$c,$d,$e);
870
871	 eval(shift(@insns));
872	 eval(shift(@insns));
873	 eval(shift(@insns));
874	 eval(shift(@insns));
875	 eval(shift(@insns));
876	 eval(shift(@insns));
877	 eval(shift(@insns));
878	&pshufb	(@X[($Xi-3)&7],@X[2]);
879	 eval(shift(@insns));
880	 eval(shift(@insns));
881	 eval(shift(@insns));
882	 eval(shift(@insns));
883	&paddd	(@X[($Xi-4)&7],@X[3]);
884	 eval(shift(@insns));
885	 eval(shift(@insns));
886	 eval(shift(@insns));
887	 eval(shift(@insns));
888	&movdqa	(&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]);	# X[]+K xfer to IALU
889	 eval(shift(@insns));
890	 eval(shift(@insns));
891	 eval(shift(@insns));
892	 eval(shift(@insns));
893	&psubd	(@X[($Xi-4)&7],@X[3]);
894
895	foreach (@insns) { eval; }
896  $Xi++;
897}
898
899sub Xtail_ssse3()
900{ use integer;
901  my $body = shift;
902  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
903  my ($a,$b,$c,$d,$e);
904
905	foreach (@insns) { eval; }
906}
907
908sub body_00_19 () {	# ((c^d)&b)^d
909	# on start @T[0]=(c^d)&b
910	return &body_20_39()	if ($rx==19);	$rx++;
911	(
912	'($a,$b,$c,$d,$e)=@V;'.
913	'&$_ror	($b,$j?7:2);',	# $b>>>2
914	'&xor	(@T[0],$d);',
915	'&mov	(@T[1],$a);',	# $b in next round
916
917	'&add	($e,&DWP(4*($j&15),"esp"));',	# X[]+K xfer
918	'&xor	($b,$c);',	# $c^$d for next round
919
920	'&$_rol	($a,5);',
921	'&add	($e,@T[0]);',
922	'&and	(@T[1],$b);',	# ($b&($c^$d)) for next round
923
924	'&xor	($b,$c);',	# restore $b
925	'&add	($e,$a);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
926	);
927}
928
929sub body_20_39 () {	# b^d^c
930	# on entry @T[0]=b^d
931	return &body_40_59()	if ($rx==39);	$rx++;
932	(
933	'($a,$b,$c,$d,$e)=@V;'.
934	'&add	($e,&DWP(4*($j&15),"esp"));',	# X[]+K xfer
935	'&xor	(@T[0],$d)	if($j==19);'.
936	'&xor	(@T[0],$c)	if($j> 19);',	# ($b^$d^$c)
937	'&mov	(@T[1],$a);',	# $b in next round
938
939	'&$_rol	($a,5);',
940	'&add	($e,@T[0]);',
941	'&xor	(@T[1],$c)	if ($j< 79);',	# $b^$d for next round
942
943	'&$_ror	($b,7);',	# $b>>>2
944	'&add	($e,$a);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
945	);
946}
947
948sub body_40_59 () {	# ((b^c)&(c^d))^c
949	# on entry @T[0]=(b^c), (c^=d)
950	$rx++;
951	(
952	'($a,$b,$c,$d,$e)=@V;'.
953	'&add	($e,&DWP(4*($j&15),"esp"));',	# X[]+K xfer
954	'&and	(@T[0],$c)	if ($j>=40);',	# (b^c)&(c^d)
955	'&xor	($c,$d)		if ($j>=40);',	# restore $c
956
957	'&$_ror	($b,7);',	# $b>>>2
958	'&mov	(@T[1],$a);',	# $b for next round
959	'&xor	(@T[0],$c);',
960
961	'&$_rol	($a,5);',
962	'&add	($e,@T[0]);',
963	'&xor	(@T[1],$c)	if ($j==59);'.
964	'&xor	(@T[1],$b)	if ($j< 59);',	# b^c for next round
965
966	'&xor	($b,$c)		if ($j< 59);',	# c^d for next round
967	'&add	($e,$a);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
968	);
969}
970######
971sub bodyx_00_19 () {	# ((c^d)&b)^d
972	# on start @T[0]=(b&c)^(~b&d), $e+=X[]+K
973	return &bodyx_20_39()	if ($rx==19);	$rx++;
974	(
975	'($a,$b,$c,$d,$e)=@V;'.
976
977	'&rorx	($b,$b,2)			if ($j==0);'.	# $b>>>2
978	'&rorx	($b,@T[1],7)			if ($j!=0);',	# $b>>>2
979	'&lea	($e,&DWP(0,$e,@T[0]));',
980	'&rorx	(@T[0],$a,5);',
981
982	'&andn	(@T[1],$a,$c);',
983	'&and	($a,$b)',
984	'&add	($d,&DWP(4*(($j+1)&15),"esp"));',	# X[]+K xfer
985
986	'&xor	(@T[1],$a)',
987	'&add	($e,@T[0]);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
988	);
989}
990
991sub bodyx_20_39 () {	# b^d^c
992	# on start $b=b^c^d
993	return &bodyx_40_59()	if ($rx==39);	$rx++;
994	(
995	'($a,$b,$c,$d,$e)=@V;'.
996
997	'&add	($e,($j==19?@T[0]:$b))',
998	'&rorx	($b,@T[1],7);',	# $b>>>2
999	'&rorx	(@T[0],$a,5);',
1000
1001	'&xor	($a,$b)				if ($j<79);',
1002	'&add	($d,&DWP(4*(($j+1)&15),"esp"))	if ($j<79);',	# X[]+K xfer
1003	'&xor	($a,$c)				if ($j<79);',
1004	'&add	($e,@T[0]);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1005	);
1006}
1007
1008sub bodyx_40_59 () {	# ((b^c)&(c^d))^c
1009	# on start $b=((b^c)&(c^d))^c
1010	return &bodyx_20_39()	if ($rx==59);	$rx++;
1011	(
1012	'($a,$b,$c,$d,$e)=@V;'.
1013
1014	'&rorx	(@T[0],$a,5)',
1015	'&lea	($e,&DWP(0,$e,$b))',
1016	'&rorx	($b,@T[1],7)',	# $b>>>2
1017	'&add	($d,&DWP(4*(($j+1)&15),"esp"))',	# X[]+K xfer
1018
1019	'&mov	(@T[1],$c)',
1020	'&xor	($a,$b)',	# b^c for next round
1021	'&xor	(@T[1],$b)',	# c^d for next round
1022
1023	'&and	($a,@T[1])',
1024	'&add	($e,@T[0])',
1025	'&xor	($a,$b)'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1026	);
1027}
1028
1029&set_label("loop",16);
1030	&Xupdate_ssse3_16_31(\&body_00_19);
1031	&Xupdate_ssse3_16_31(\&body_00_19);
1032	&Xupdate_ssse3_16_31(\&body_00_19);
1033	&Xupdate_ssse3_16_31(\&body_00_19);
1034	&Xupdate_ssse3_32_79(\&body_00_19);
1035	&Xupdate_ssse3_32_79(\&body_20_39);
1036	&Xupdate_ssse3_32_79(\&body_20_39);
1037	&Xupdate_ssse3_32_79(\&body_20_39);
1038	&Xupdate_ssse3_32_79(\&body_20_39);
1039	&Xupdate_ssse3_32_79(\&body_20_39);
1040	&Xupdate_ssse3_32_79(\&body_40_59);
1041	&Xupdate_ssse3_32_79(\&body_40_59);
1042	&Xupdate_ssse3_32_79(\&body_40_59);
1043	&Xupdate_ssse3_32_79(\&body_40_59);
1044	&Xupdate_ssse3_32_79(\&body_40_59);
1045	&Xupdate_ssse3_32_79(\&body_20_39);
1046	&Xuplast_ssse3_80(\&body_20_39);	# can jump to "done"
1047
1048				$saved_j=$j; @saved_V=@V;
1049
1050	&Xloop_ssse3(\&body_20_39);
1051	&Xloop_ssse3(\&body_20_39);
1052	&Xloop_ssse3(\&body_20_39);
1053
1054	&mov	(@T[1],&DWP(192,"esp"));	# update context
1055	&add	($A,&DWP(0,@T[1]));
1056	&add	(@T[0],&DWP(4,@T[1]));		# $b
1057	&add	($C,&DWP(8,@T[1]));
1058	&mov	(&DWP(0,@T[1]),$A);
1059	&add	($D,&DWP(12,@T[1]));
1060	&mov	(&DWP(4,@T[1]),@T[0]);
1061	&add	($E,&DWP(16,@T[1]));
1062	&mov	(&DWP(8,@T[1]),$C);
1063	&mov	($B,$C);
1064	&mov	(&DWP(12,@T[1]),$D);
1065	&xor	($B,$D);
1066	&mov	(&DWP(16,@T[1]),$E);
1067	&mov	(@T[1],@T[0]);
1068	&pshufd	(@X[0],@X[-4&7],0xee);		# was &movdqa	(@X[0],@X[-3&7]);
1069	&and	(@T[0],$B);
1070	&mov	($B,$T[1]);
1071
1072	&jmp	(&label("loop"));
1073
1074&set_label("done",16);		$j=$saved_j; @V=@saved_V;
1075
1076	&Xtail_ssse3(\&body_20_39);
1077	&Xtail_ssse3(\&body_20_39);
1078	&Xtail_ssse3(\&body_20_39);
1079
1080	&mov	(@T[1],&DWP(192,"esp"));	# update context
1081	&add	($A,&DWP(0,@T[1]));
1082	&mov	("esp",&DWP(192+12,"esp"));	# restore %esp
1083	&add	(@T[0],&DWP(4,@T[1]));		# $b
1084	&add	($C,&DWP(8,@T[1]));
1085	&mov	(&DWP(0,@T[1]),$A);
1086	&add	($D,&DWP(12,@T[1]));
1087	&mov	(&DWP(4,@T[1]),@T[0]);
1088	&add	($E,&DWP(16,@T[1]));
1089	&mov	(&DWP(8,@T[1]),$C);
1090	&mov	(&DWP(12,@T[1]),$D);
1091	&mov	(&DWP(16,@T[1]),$E);
1092
1093&function_end("_sha1_block_data_order_ssse3");
1094
1095$rx=0;	# reset
1096
1097if ($ymm) {
1098my $Xi=4;			# 4xSIMD Xupdate round, start pre-seeded
1099my @X=map("xmm$_",(4..7,0..3));	# pre-seeded for $Xi=4
1100my @V=($A,$B,$C,$D,$E);
1101my $j=0;			# hash round
1102my @T=($T,$tmp1);
1103my $inp;
1104
1105my $_rol=sub { &shld(@_[0],@_) };
1106my $_ror=sub { &shrd(@_[0],@_) };
1107
1108&function_begin("_sha1_block_data_order_avx");
1109	&call	(&label("pic_point"));	# make it PIC!
1110	&set_label("pic_point");
1111	&blindpop($tmp1);
1112	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
1113&set_label("avx_shortcut");
1114	&vzeroall();
1115
1116	&vmovdqa(@X[3],&QWP(0,$tmp1));		# K_00_19
1117	&vmovdqa(@X[4],&QWP(16,$tmp1));		# K_20_39
1118	&vmovdqa(@X[5],&QWP(32,$tmp1));		# K_40_59
1119	&vmovdqa(@X[6],&QWP(48,$tmp1));		# K_60_79
1120	&vmovdqa(@X[2],&QWP(64,$tmp1));		# pbswap mask
1121
1122	&mov	($E,&wparam(0));		# load argument block
1123	&mov	($inp=@T[1],&wparam(1));
1124	&mov	($D,&wparam(2));
1125	&mov	(@T[0],"esp");
1126
1127	# stack frame layout
1128	#
1129	# +0	X[0]+K	X[1]+K	X[2]+K	X[3]+K	# XMM->IALU xfer area
1130	#	X[4]+K	X[5]+K	X[6]+K	X[7]+K
1131	#	X[8]+K	X[9]+K	X[10]+K	X[11]+K
1132	#	X[12]+K	X[13]+K	X[14]+K	X[15]+K
1133	#
1134	# +64	X[0]	X[1]	X[2]	X[3]	# XMM->XMM backtrace area
1135	#	X[4]	X[5]	X[6]	X[7]
1136	#	X[8]	X[9]	X[10]	X[11]	# even borrowed for K_00_19
1137	#
1138	# +112	K_20_39	K_20_39	K_20_39	K_20_39	# constants
1139	#	K_40_59	K_40_59	K_40_59	K_40_59
1140	#	K_60_79	K_60_79	K_60_79	K_60_79
1141	#	K_00_19	K_00_19	K_00_19	K_00_19
1142	#	pbswap mask
1143	#
1144	# +192	ctx				# argument block
1145	# +196	inp
1146	# +200	end
1147	# +204	esp
1148	&sub	("esp",208);
1149	&and	("esp",-64);
1150
1151	&vmovdqa(&QWP(112+0,"esp"),@X[4]);	# copy constants
1152	&vmovdqa(&QWP(112+16,"esp"),@X[5]);
1153	&vmovdqa(&QWP(112+32,"esp"),@X[6]);
1154	&shl	($D,6);				# len*64
1155	&vmovdqa(&QWP(112+48,"esp"),@X[3]);
1156	&add	($D,$inp);			# end of input
1157	&vmovdqa(&QWP(112+64,"esp"),@X[2]);
1158	&add	($inp,64);
1159	&mov	(&DWP(192+0,"esp"),$E);		# save argument block
1160	&mov	(&DWP(192+4,"esp"),$inp);
1161	&mov	(&DWP(192+8,"esp"),$D);
1162	&mov	(&DWP(192+12,"esp"),@T[0]);	# save original %esp
1163
1164	&mov	($A,&DWP(0,$E));		# load context
1165	&mov	($B,&DWP(4,$E));
1166	&mov	($C,&DWP(8,$E));
1167	&mov	($D,&DWP(12,$E));
1168	&mov	($E,&DWP(16,$E));
1169	&mov	(@T[0],$B);			# magic seed
1170
1171	&vmovdqu(@X[-4&7],&QWP(-64,$inp));	# load input to %xmm[0-3]
1172	&vmovdqu(@X[-3&7],&QWP(-48,$inp));
1173	&vmovdqu(@X[-2&7],&QWP(-32,$inp));
1174	&vmovdqu(@X[-1&7],&QWP(-16,$inp));
1175	&vpshufb(@X[-4&7],@X[-4&7],@X[2]);	# byte swap
1176	&vpshufb(@X[-3&7],@X[-3&7],@X[2]);
1177	&vpshufb(@X[-2&7],@X[-2&7],@X[2]);
1178	&vmovdqa(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
1179	&vpshufb(@X[-1&7],@X[-1&7],@X[2]);
1180	&vpaddd	(@X[0],@X[-4&7],@X[3]);		# add K_00_19
1181	&vpaddd	(@X[1],@X[-3&7],@X[3]);
1182	&vpaddd	(@X[2],@X[-2&7],@X[3]);
1183	&vmovdqa(&QWP(0,"esp"),@X[0]);		# X[]+K xfer to IALU
1184	&mov	(@T[1],$C);
1185	&vmovdqa(&QWP(0+16,"esp"),@X[1]);
1186	&xor	(@T[1],$D);
1187	&vmovdqa(&QWP(0+32,"esp"),@X[2]);
1188	&and	(@T[0],@T[1]);
1189	&jmp	(&label("loop"));
1190
1191sub Xupdate_avx_16_31()		# recall that $Xi starts wtih 4
1192{ use integer;
1193  my $body = shift;
1194  my @insns = (&$body,&$body,&$body,&$body);	# 40 instructions
1195  my ($a,$b,$c,$d,$e);
1196
1197	 eval(shift(@insns));
1198	 eval(shift(@insns));
1199	&vpalignr(@X[0],@X[-3&7],@X[-4&7],8);	# compose "X[-14]" in "X[0]"
1200	 eval(shift(@insns));
1201	 eval(shift(@insns));
1202
1203	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
1204	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
1205	 eval(shift(@insns));
1206	 eval(shift(@insns));
1207	&vpsrldq(@X[2],@X[-1&7],4);		# "X[-3]", 3 dwords
1208	 eval(shift(@insns));
1209	 eval(shift(@insns));
1210	&vpxor	(@X[0],@X[0],@X[-4&7]);		# "X[0]"^="X[-16]"
1211	 eval(shift(@insns));
1212	 eval(shift(@insns));
1213
1214	&vpxor	(@X[2],@X[2],@X[-2&7]);		# "X[-3]"^"X[-8]"
1215	 eval(shift(@insns));
1216	 eval(shift(@insns));
1217	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
1218	 eval(shift(@insns));
1219	 eval(shift(@insns));
1220
1221	&vpxor	(@X[0],@X[0],@X[2]);		# "X[0]"^="X[-3]"^"X[-8]"
1222	 eval(shift(@insns));
1223	 eval(shift(@insns));
1224	 eval(shift(@insns));
1225	 eval(shift(@insns));
1226
1227	&vpsrld	(@X[2],@X[0],31);
1228	 eval(shift(@insns));
1229	 eval(shift(@insns));
1230	 eval(shift(@insns));
1231	 eval(shift(@insns));
1232
1233	&vpslldq(@X[4],@X[0],12);		# "X[0]"<<96, extract one dword
1234	&vpaddd	(@X[0],@X[0],@X[0]);
1235	 eval(shift(@insns));
1236	 eval(shift(@insns));
1237	 eval(shift(@insns));
1238	 eval(shift(@insns));
1239
1240	&vpsrld	(@X[3],@X[4],30);
1241	&vpor	(@X[0],@X[0],@X[2]);		# "X[0]"<<<=1
1242	 eval(shift(@insns));
1243	 eval(shift(@insns));
1244	 eval(shift(@insns));
1245	 eval(shift(@insns));
1246
1247	&vpslld	(@X[4],@X[4],2);
1248	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
1249	 eval(shift(@insns));
1250	 eval(shift(@insns));
1251	&vpxor	(@X[0],@X[0],@X[3]);
1252	 eval(shift(@insns));
1253	 eval(shift(@insns));
1254	 eval(shift(@insns));
1255	 eval(shift(@insns));
1256
1257	&vpxor	(@X[0],@X[0],@X[4]);		# "X[0]"^=("X[0]"<<96)<<<2
1258	 eval(shift(@insns));
1259	 eval(shift(@insns));
1260	  &vmovdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
1261	 eval(shift(@insns));
1262	 eval(shift(@insns));
1263
1264	 foreach (@insns) { eval; }	# remaining instructions [if any]
1265
1266  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
1267}
1268
1269sub Xupdate_avx_32_79()
1270{ use integer;
1271  my $body = shift;
1272  my @insns = (&$body,&$body,&$body,&$body);	# 32 to 44 instructions
1273  my ($a,$b,$c,$d,$e);
1274
1275	&vpalignr(@X[2],@X[-1&7],@X[-2&7],8);	# compose "X[-6]"
1276	&vpxor	(@X[0],@X[0],@X[-4&7]);	# "X[0]"="X[-32]"^"X[-16]"
1277	 eval(shift(@insns));		# body_20_39
1278	 eval(shift(@insns));
1279	 eval(shift(@insns));
1280	 eval(shift(@insns));		# rol
1281
1282	&vpxor	(@X[0],@X[0],@X[-7&7]);	# "X[0]"^="X[-28]"
1283	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
1284	 eval(shift(@insns));
1285	 eval(shift(@insns));
1286	 if ($Xi%5) {
1287	  &vmovdqa	(@X[4],@X[3]);	# "perpetuate" K_XX_XX...
1288	 } else {			# ... or load next one
1289	  &vmovdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
1290	 }
1291	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
1292	 eval(shift(@insns));		# ror
1293	 eval(shift(@insns));
1294
1295	&vpxor	(@X[0],@X[0],@X[2]);		# "X[0]"^="X[-6]"
1296	 eval(shift(@insns));		# body_20_39
1297	 eval(shift(@insns));
1298	 eval(shift(@insns));
1299	 eval(shift(@insns));		# rol
1300
1301	&vpsrld	(@X[2],@X[0],30);
1302	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
1303	 eval(shift(@insns));
1304	 eval(shift(@insns));
1305	 eval(shift(@insns));		# ror
1306	 eval(shift(@insns));
1307
1308	&vpslld	(@X[0],@X[0],2);
1309	 eval(shift(@insns));		# body_20_39
1310	 eval(shift(@insns));
1311	 eval(shift(@insns));
1312	 eval(shift(@insns));		# rol
1313	 eval(shift(@insns));
1314	 eval(shift(@insns));
1315	 eval(shift(@insns));		# ror
1316	 eval(shift(@insns));
1317
1318	&vpor	(@X[0],@X[0],@X[2]);	# "X[0]"<<<=2
1319	 eval(shift(@insns));		# body_20_39
1320	 eval(shift(@insns));
1321	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
1322	 eval(shift(@insns));
1323	 eval(shift(@insns));		# rol
1324	 eval(shift(@insns));
1325	 eval(shift(@insns));
1326	 eval(shift(@insns));		# ror
1327	 eval(shift(@insns));
1328
1329	 foreach (@insns) { eval; }	# remaining instructions
1330
1331  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
1332}
1333
1334sub Xuplast_avx_80()
1335{ use integer;
1336  my $body = shift;
1337  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1338  my ($a,$b,$c,$d,$e);
1339
1340	 eval(shift(@insns));
1341	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
1342	 eval(shift(@insns));
1343	 eval(shift(@insns));
1344	 eval(shift(@insns));
1345	 eval(shift(@insns));
1346
1347	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU
1348
1349	 foreach (@insns) { eval; }		# remaining instructions
1350
1351	&mov	($inp=@T[1],&DWP(192+4,"esp"));
1352	&cmp	($inp,&DWP(192+8,"esp"));
1353	&je	(&label("done"));
1354
1355	&vmovdqa(@X[3],&QWP(112+48,"esp"));	# K_00_19
1356	&vmovdqa(@X[2],&QWP(112+64,"esp"));	# pbswap mask
1357	&vmovdqu(@X[-4&7],&QWP(0,$inp));	# load input
1358	&vmovdqu(@X[-3&7],&QWP(16,$inp));
1359	&vmovdqu(@X[-2&7],&QWP(32,$inp));
1360	&vmovdqu(@X[-1&7],&QWP(48,$inp));
1361	&add	($inp,64);
1362	&vpshufb(@X[-4&7],@X[-4&7],@X[2]);		# byte swap
1363	&mov	(&DWP(192+4,"esp"),$inp);
1364	&vmovdqa(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
1365
1366  $Xi=0;
1367}
1368
1369sub Xloop_avx()
1370{ use integer;
1371  my $body = shift;
1372  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1373  my ($a,$b,$c,$d,$e);
1374
1375	 eval(shift(@insns));
1376	 eval(shift(@insns));
1377	&vpshufb	(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1378	 eval(shift(@insns));
1379	 eval(shift(@insns));
1380	&vpaddd	(@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
1381	 eval(shift(@insns));
1382	 eval(shift(@insns));
1383	 eval(shift(@insns));
1384	 eval(shift(@insns));
1385	&vmovdqa	(&QWP(0+16*$Xi,"esp"),@X[$Xi&7]);	# X[]+K xfer to IALU
1386	 eval(shift(@insns));
1387	 eval(shift(@insns));
1388
1389	foreach (@insns) { eval; }
1390  $Xi++;
1391}
1392
1393sub Xtail_avx()
1394{ use integer;
1395  my $body = shift;
1396  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1397  my ($a,$b,$c,$d,$e);
1398
1399	foreach (@insns) { eval; }
1400}
1401
1402&set_label("loop",16);
1403	&Xupdate_avx_16_31(\&body_00_19);
1404	&Xupdate_avx_16_31(\&body_00_19);
1405	&Xupdate_avx_16_31(\&body_00_19);
1406	&Xupdate_avx_16_31(\&body_00_19);
1407	&Xupdate_avx_32_79(\&body_00_19);
1408	&Xupdate_avx_32_79(\&body_20_39);
1409	&Xupdate_avx_32_79(\&body_20_39);
1410	&Xupdate_avx_32_79(\&body_20_39);
1411	&Xupdate_avx_32_79(\&body_20_39);
1412	&Xupdate_avx_32_79(\&body_20_39);
1413	&Xupdate_avx_32_79(\&body_40_59);
1414	&Xupdate_avx_32_79(\&body_40_59);
1415	&Xupdate_avx_32_79(\&body_40_59);
1416	&Xupdate_avx_32_79(\&body_40_59);
1417	&Xupdate_avx_32_79(\&body_40_59);
1418	&Xupdate_avx_32_79(\&body_20_39);
1419	&Xuplast_avx_80(\&body_20_39);	# can jump to "done"
1420
1421				$saved_j=$j; @saved_V=@V;
1422
1423	&Xloop_avx(\&body_20_39);
1424	&Xloop_avx(\&body_20_39);
1425	&Xloop_avx(\&body_20_39);
1426
1427	&mov	(@T[1],&DWP(192,"esp"));	# update context
1428	&add	($A,&DWP(0,@T[1]));
1429	&add	(@T[0],&DWP(4,@T[1]));		# $b
1430	&add	($C,&DWP(8,@T[1]));
1431	&mov	(&DWP(0,@T[1]),$A);
1432	&add	($D,&DWP(12,@T[1]));
1433	&mov	(&DWP(4,@T[1]),@T[0]);
1434	&add	($E,&DWP(16,@T[1]));
1435	&mov	($B,$C);
1436	&mov	(&DWP(8,@T[1]),$C);
1437	&xor	($B,$D);
1438	&mov	(&DWP(12,@T[1]),$D);
1439	&mov	(&DWP(16,@T[1]),$E);
1440	&mov	(@T[1],@T[0]);
1441	&and	(@T[0],$B);
1442	&mov	($B,@T[1]);
1443
1444	&jmp	(&label("loop"));
1445
1446&set_label("done",16);		$j=$saved_j; @V=@saved_V;
1447
1448	&Xtail_avx(\&body_20_39);
1449	&Xtail_avx(\&body_20_39);
1450	&Xtail_avx(\&body_20_39);
1451
1452	&vzeroall();
1453
1454	&mov	(@T[1],&DWP(192,"esp"));	# update context
1455	&add	($A,&DWP(0,@T[1]));
1456	&mov	("esp",&DWP(192+12,"esp"));	# restore %esp
1457	&add	(@T[0],&DWP(4,@T[1]));		# $b
1458	&add	($C,&DWP(8,@T[1]));
1459	&mov	(&DWP(0,@T[1]),$A);
1460	&add	($D,&DWP(12,@T[1]));
1461	&mov	(&DWP(4,@T[1]),@T[0]);
1462	&add	($E,&DWP(16,@T[1]));
1463	&mov	(&DWP(8,@T[1]),$C);
1464	&mov	(&DWP(12,@T[1]),$D);
1465	&mov	(&DWP(16,@T[1]),$E);
1466&function_end("_sha1_block_data_order_avx");
1467}
1468&set_label("K_XX_XX",64);
1469&data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999);	# K_00_19
1470&data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1);	# K_20_39
1471&data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc);	# K_40_59
1472&data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6);	# K_60_79
1473&data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);	# pbswap mask
1474&data_byte(0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0);
1475}
1476&asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
1477
1478&asm_finish();
1479
1480close STDOUT;
1481